Hyperspectral image classification based on deep stacking network

Hyperspectral image (HIS) classification is a hot topic in remote sensing community and most of the existing methods extract the features of original Hyperspectral data using shallow layer networks such as neural network (NN) and support vector machine (SVM). As deep learning recently achieves great success in machine learning and pattern recognition area for its ability in deep feature extraction and representations, two deep networks i.e. deep convolutional network (DCN) and deep belief network (DBN) have been used for hyperspectral image classification and better results have been achieved. Differing from those deep networks for HSI classification, in this paper, we propose a new method for hyperspectral image classification based on deep stacking network (DSN), which owns advantages to other deep models for its simplicity when processing in batch-mode learning - not requiring stochastic gradient descent that other DNNs require. The feature extraction is gradually obtained by employing nonlinear activation function on the hidden layer nodes of each module, which is different from those DSNs that usually use linear weights between the hidden layer and the output layer. Experimental results on AVIRIS hyperspectral images show that the proposed method achieves improved classification performance when compared with that via SVM and NN methods.